Fisk-Ai

Today I released version 0.0.3 of the Fisk AI Project.

This is a release focussed mainly on internal changes to prepare it for the longer term goal of being a harness I can use to solve real Operations problems using LLMs - safely, focussed, guardrailed, audited and controlled. All running locally if needed.

I’ve made every sub system pluggable via Adapters and Providers:

• The LLM protocol can be replaced
• The Memory store can be replaced
• The Session store can be replaced
• The Vector store can be replaced
• The A2A transport can be replaced
• The Human in the Loop callout methods can be replaced
• Agents can be suspended entirely, complete compute resources freed and resumed later while waiting for HITL feedback etc
• Tools can be added programmatically via Go functions
• The entire agent loop is thread safe and can run in Go routines

Using these capabilities we added an, optional, NATS JetStream backend for Sessions and Memory allowing these to be shared across infrastructure.

I’ve recorded a video that introduces the release and discuss the goals

Next I’ll work on integrating Choria Async Jobs as a way to get work into the system.

Some other links of note:

• Release Notes
• Project Roadmap
• A deep study of the code based generated by AI Code Map
• The Claude Plugin that generated the Code Map
• An interesting video about Domain Specific Agents very close to what we are working on here

Today I would like to introduce an AI Harness I’ve been working on for the last 2 or 3 months called Fisk AI.

There are many fantastic, all powerful, all featureful AI Harnesses out there like Claude Code, Pi, Hermess, Codex, Open Code, and many more. At the same time there’s been a movement toward creating harnesses that solve just one problem and do so well and safely.

Fisk AI is a Harness framework that allows you to build Specialized AI Agents with just some YAML files.

• Agentic loop in TUI and Shell form
• Easiest possible way to create tools reusing the CLI tools you already have and, optionally, expose those over MCP
• Built-in Memory System
• Built-in Knowledge base (RAG)

Fisk AI can be used entirely locally on your laptop with locally hosted models or can use any AI host that supports the Anthropic API (most do).

Data safety and tool safety is at the forefront, big harnesses have tons of capabilities and love to try and please you. It is not uncommon to see Claude run half-page long Shell scripts, Python scripts, jq scripts, or combinations of these in rapid succession. Mistakes are inevitable and so are non-deterministic outcomes. Claude ran Bashism on my ZSH and did rm -rf /by acccident.

There is no reason to have this power around and a LLM that’s keen to go the extra mile when you have a specific problem to solve, repeatedly and reliably. This is where Specialized AI comes in. Give the AI just the tools and guidance it needs, no generic shell access or ability to run arbitrary code. That’s what Fisk AI is for. Better, safer, and more reliable outcomes for tasks you wish to run regularly with LLM assistance.

Down the line there is a broader goal around enterprise AI and infrastructure Ops. For now I am focussing on getting the libraries and behaviours right, later the libraries that built Fisk AI will become the basis for infrastructure tools onto of the Choria protocol which will bring strong Identity, Authentication, Authorization, and Auditing to Agent-2-Agent communications.

Read the full entry for examples and more details.